The medicine was already in the country.
Aarogya Grid forecasts stock-outs across India’s primary health network, then finds the surplus sitting in a district nearby and moves it before the shelf goes empty. Not more procurement — better circulation of what has already been bought.
Demand is forecast by TimesFM 2.0 through BigQuery AI.FORECAST, in the class where a held-out backtest says it wins. Gemini reads a paper register or a spoken Hindi report, and answers questions over the computed state with its tool calls on screen. Neither model invents a number.
Try this in 60 seconds
- 1Open the console — 12,010 facilities, 3.5 L stock positions, 16,772 of them critical today.
- 2Scroll one screen to Ask the grid and press “Where is it worst tonight?” — in English, Hindi or Hinglish. The audit trail beside the answer lists every tool that ran.
- 3Open any district and find a dispatch order that crosses a boundary. It names a batch and an expiry date — and Approve is disabled until the donor district countersigns.
- 34.3 L
- units of shortfall averted
- 2,348
- inter-district corridors
- 121.12 Cr
- people in catchment
What the plan actually does
34,39,003
units of medicine that were forecast to run out, and now do not — filled from stock that already existed somewhere else in the network.
23,070
dispatches
individual facility-to-facility orders in the plan
9,421
vehicle trips
orders sharing a route share a vehicle
6,415
cross a district line
carrying 15,930 orders
11,279
ride along
admitted only because a vehicle was already going
The part most decks leave out
It does not pay for itself in cash.
Redistribution spends more moving stock than it recovers in averted expiry. That is not a rounding error to be presented away — it is the actual shape of the intervention, and any figure that hid it would fall apart the moment someone opened the console.
So the case is put the other way round. Rather than claiming a return, the plan states the price at which the return exists: it breaks even when one averted unit of unmet demand is worth ₹3.11. Whether a dose of a Vital medicine reaching a patient is worth that is a policy judgement, not an engineering one.
Plan economics · national
Pricing a route once instead of once per drug is what pays for crossing a boundary at all — consolidation takes transport from ₹2.77 Cr to ₹1.26 Cr, a saving of ₹1.51 Cr.
How it works
Four stages, one shared allocation state.
01
Simulate
A year of stock ledger for 3.5 L facility × drug positions — receipts, consumption, batches and expiry — seeded so any figure can be regenerated exactly.
02
Forecast
Demand and stock-out probability per position. 16,772 come back critical; 20,511 are already at zero on the shelf.
03
Plan
Match surplus to shortfall against a benefit/cost gate. A donor batch can be promised only once, so the whole run shares one allocation state rather than solving districts independently.
04
Dispatch
Consolidate orders onto shared vehicles and let them cross district lines — 6,415 of 9,421 trips do, on 2,348 corridors.
Because planning shares one allocation state, districts are not independent and the plan is order-dependent — deterministic, not symmetric. It parallelises where clusters are disjoint, which on this build is 10 rounds rather than 769 tasks. Simulation and forecasting remain embarrassingly parallel.
Depth and reach
Built at national scale, not demoed on one district.
- 769
- districtsacross 36 states
- 12,010
- facilitiesDH, CHC, PHC and sub-centre
- 3.5 L
- stock positionsfacility × drug pairs tracked
- 121.12 Cr
- people coveredmodelled catchment population
- 2,348
- corridors284 cross a state line
- 763
- districts on a corridorof 769 — the rest are self-sufficient
- 361.4s
- to build the countryone machine, one batch run
- ₹9.47 Cr
- net benefitunmet demand valued at the VED policy multiple · cash alone −₹1.07 Cr
Provenance
What is real here, and what is not.
No public PHC inventory feed exists in India. The facility layer is generated by a seeded simulator, parameterised from IPHS norms and published epidemiological seasonality — fitted to those norms, not to observed consumption. Saying so plainly is cheaper than being caught not saying it.
Real
- Districts, state LGD/Census codes, and coordinates
- Kerala’s IDSP daily disease bulletins, behind the observed warning signals
- IPHS facility norms, catchment norms and bed strength
- IPHS staffing establishment by tier and cadre
- Drug catalogue, VED classification, cold-chain flags
- Every model, forecast and optimisation in the system
Simulated
- Individual facilities, their names and catchments
- Stock positions, batches and consumption ledgers
- Bed occupancy, staff vacancy and daily attendance
- District supply reliability and allocation behaviour
One caveat the console repeats and this page will not bury: the workforce layer and the stock layer are correlated by construction. Remoteness is derived from the same synthetic district reliability parameter that drives supply, so the fact that badly-supplied districts are also badly-staffed here is an assumption in the model, not a finding from it — 918 stock-holding facilities have no pharmacist in position, against 18.7% vacancy and 16.7% absence, and that is why the stock board carries an error bar.
Open it and go looking for the seams.
Every figure on this page is read from the same shipped snapshot the console renders. Drill into any district, follow any corridor, and check the arithmetic.